If you've ever tried to track individual cells under a microscope, you understand suffering on a level that most people don't. Cells divide, creating two problems where before you had one. They move around. They look almost identical to every other cell nearby. Now imagine doing this for millions of cells over hours of video footage, and you've got a computational nightmare that has driven biologists to drink for decades.

A new machine learning tool called CELLECT, published in Nature Methods, actually handles this problem well. It tracks large-scale cellular behaviors in real-time with impressive accuracy. And unlike most AI tools that need to be completely retrained every time you change anything about your experiment, CELLECT can generalize across different imaging systems and even different species. It's the rare AI that learns principles instead of just memorizing patterns.

Why Cell Tracking Has Been Such a Nightmare

Understanding how cells behave is fundamental to biology. Immune responses, neural activity, cancer progression, embryonic development. All of these require knowing what individual cells are doing over time. Are they dividing? Dying? Moving? Changing shape? Eating bacteria?

The problem is that getting answers to these questions requires tracking individual cells through time-lapse imaging, and traditional approaches have been brutal. You need massive amounts of labeled training data for each specific application. Changed your microscope settings? Time to retrain your model. Switched to a different cell type? Retrain. Looking at a different organism? You know where this is going.

This retraining requirement has meant that even labs with expensive AI tools often end up doing much of their cell tracking semi-manually, which is slow, error-prone, and makes researchers question their career choices.

Teaching AI to Actually Understand Cells

CELLECT uses something called contrastive embedding learning, which is a fancy way of saying it learns to recognize fundamental features of cells and their movements that don't change just because you switched microscopes or species.

Instead of learning "this specific blob of pixels is a B cell in this specific imaging setup," CELLECT learns something more like "this is what it looks like when a cell is a cell, and this is what cell movement patterns look like, and this is how cells relate to each other across time frames." These representations transfer across contexts because they capture what's actually universal about cells rather than what's incidental to a particular experimental setup.

The result is that you can pretrain CELLECT once on one public dataset and then use it on completely different tasks without retraining. That's huge.

One Ring to Rule Them All (But For Cells)

The researchers put CELLECT through a gauntlet of wildly different applications. Real-time 3D tracking of B cells during germinal center formation in mouse lymph nodes. Quantifying how cells interact with bacteria in mouse spleens. Extracting neural signals from worms that were wiggling around (which creates motion artifacts that normally ruin everything).

One model handled all of these. Not three different models that each needed separate training. One model, pretrained on one public dataset, that generalized to this whole range of challenges. That's not just convenient. That's a fundamental shift in how imaging-based research can work.

The Wiggling Worm Problem (Solved)

There's a specific part of this that should make neuroscientists sit up and take notice: "high-fidelity extraction of neural signals during strong nonrigid motions."

Here's the issue. When you're imaging neural activity in behaving animals, the animals have the inconsiderate habit of moving around. The brain itself shifts position. Individual neurons drift relative to your imaging plane. Your beautiful recordings become corrupted by motion artifacts, and figuring out which neuron is which across time becomes a mess.

CELLECT's ability to maintain accurate tracking despite this kind of movement could dramatically improve the reliability of neural recordings in behaving animals. Instead of restricting studies to paralyzed or head-fixed animals, researchers could potentially get clean data from more naturalistic situations. That opens up whole categories of experiments that were previously impractical.

What This Means for Biology Labs

This is part of a broader trend in AI tools: models that actually generalize instead of needing to be retrained for every slight variation in experimental conditions. For research labs drowning in imaging data, this isn't just a convenience. It's potentially transformative.

The old model was: get imaging data, spend weeks training a custom AI for your specific setup, pray that it works, and start over whenever anything changes. The new model is: use a pretrained tool that already understands what cells look like and how they behave, and get on with the actual science.

Cell tracking used to be a bottleneck that limited what questions researchers could practically answer. Tools like CELLECT might remove that bottleneck, which means more questions get asked and answered. Sometimes the biggest advances aren't in what we study but in how efficiently we can study it.

Reference: Zhou H, et al. (2025). CELLECT: contrastive embedding learning for large-scale efficient cell tracking. Nature Methods. doi: 10.1038/s41592-025-02886-x | PMID: 41116016

Disclaimer: The image accompanying this article is for illustrative purposes only and does not depict actual experimental results, data, or biological mechanisms.